2017 fast conference...2017 fast conference la plata, argentina november, 2017 va smalltalk 9:...

Copyright © 2017 Instantiations, Inc.

2017 FAST Conference La Plata, Argentina

November, 2017

VA Smalltalk 9: Exploring the next-gen LLVM-based virtual

machine Alexander Mitin

Senior Software Engineer

Instantiations, Inc.


The New VM Runtime

• 64-bit & 32-bit

• Written from scratch

• Compatibility & Performance

• LLVM is a core tech


Traditional Way

• Written in C or C++

• Interpreter is a huge case statement

• Or a huge function with lots of labels

• Lots of overhead

• Support for JIT is complicated


Options?

• Continue to develop using legacy tools?

• Use assembly language?

• Develop a special language for building a VM?


Let’s C

A typical C language compiler has these stages:

• Preprocessor

• Code parsing

• Building some intermediate representation

• Optimizations

• Machine code generation

• Output to an object file


LLVM Can Do it

LLVM is a compiler infrastructure

• Advanced IR

• Object file generation

• Open-source with big community

• Capable for JIT


Instantiations Way

• Written in LLVM IR

• Every bytecode handler is a function

• Fine-tuned using LLVM intrinsics and attrs

• Minimized overhead

• Transparent interpreter-JIT transitions


A bytecode

01 I_32 BCpushMagicB(VMContext* vmStruct, oop** sp, I_8* pc)

02 {

03 I_PTR operand = (I_PTR)*pc++; // load a bytecode operand

04 *--sp = (oop*)operand; // store it to a stack slot and advance sp

05 I_8 nextBC = *pc++; // get next bytecode

06 BCFuncType* nextHandler = bytecodeTable[nextBC];// get next bytecode handler function

07 return nextHandler(vmStruct, sp, pc) // call it

08 }


LLVM Crash Course

• Modules & Functions

• Basic Blocks

• SSA form


A bytecode

01 define internal cc18 i32 @BCpushMagicB(%VMContext* inreg %vmStruct, %oop** inreg %sp, i8* inreg %pc) #1 align 8

02 {

03 %1 = load i8, i8* %pc, align 1

04 %2 = sext i8 %1 to i64

05 %3 = getelementptr inbounds i8, i8* %pc, i64 1

06 %4 = inttoptr i64 %2 to %oop*

07 %5 = getelementptr %oop*, %oop** %sp, i64 -1

08 store %oop* %4, %oop** %5

09 %6 = load i8, i8* %3

10 %7 = zext i8 %6 to i64

11 %8 = getelementptr inbounds i8, i8* %3, i64 1

12 %9 = getelementptr inbounds @bytecodeTable, i64 0, i64 %7

13 %10 = load BCFuncType*, BCFuncType** %9, align 8

14 %11 = tail call cc18 i32 %10(%VMContext* inreg %vmStruct, %oop** inreg %5, i8* inreg %8)

15 ret i32 %11

16 }


A bytecode

01 void EsVMBytecodes::pushMagic(IRBuilder<>& b, Value* sp, Value* pc, Value* bp, EsBCOpType bcot)

02 {

03 Value *operand;

04 Value *currentPC = getNext(b, pc, bcot, operand);

05 Value *currentSP = stackPush(b, sp, operand);

06 executeNextBytecode(b, currentSP, currentPC, bp);

07 }


A bytecode

00 BCpushMagicB:

01 movsbq (%rsi), %rax ; load bytecode operand from pc

02 movq %rax, -8(%r15) ; store it into a stack slot

03 addq $-8, %r15 ; advance stack pointer

04 movzbl 1(%rsi), %eax ; get next bytecode from pc + 1

05 addq $2, %rsi ; advance pc by 2

06 movabsq $bytecodeTable, %rcx

07 rex64 jmpq *(%rcx,%rax,8) ; indirect jump to the next bytecode handler


The Idea

• LLVM tail-call elimination optimization

• Keep performance-critical data in CPU regs


Method Exec: Interpreted

// a pseudo-code

I_32 methodExec(VMContext*, I_PTR arg1, I_PTR arg2)

{

AllocateStackFrame(ftInterpreted);

return ExecuteNextBytecode(arg2); // tail-call

}


Method Exec: Native

// a pseudo-code

I_32 methodExec(VMContext*, I_PTR arg1, I_PTR arg2)

{

AllocateStackFrame(ftNative);

// execute JIT-generated native code in place

}


Memory Manager

• Written in C from scratch

• So far the same algorithm as it was in v.8

• It’s ready for modifications


Build System

• CMake based

• MinGW64/gcc compiler

• A special app with LLVM IR

• MinGW64/gcc linker


Future Work

• Unicode support

• Alternative memory manager with parallel GC

• Alternative JIT engine (OMR?)

• Make use of CPU’s stack pointer register

• Compressed OOPs


Thank you for your attention

Questions?

Contacts:

E-mail: [email protected]

GitHub: amitin

mailto:[email protected]

mailto:[email protected]

2017 fast conference...2017 fast conference la plata, argentina november, 2017 va smalltalk 9:...

Documents